The champion (probabilityEstimator -> p_win) keeps serving and grading users,
completely unchanged. The challenger is a second probability computed from the
same inputs at the same instant, landing on the same ledger row so it joins to
the same outcome and the same close. Identical conditions, one difference —
the only clean A/B.
NOTHING IS CLAIMED. Running a challenger is honest beta; asserting it is better
before the settled ledger says so is not. Promotion stays a later decision gated
on Brier and calibration over sufficient segmented volume.
INTERPRETABLE, NOT A RE-ESTIMATION. The challenger is the champion's probability
adjusted by the Layer-2 axes, applied in log-odds space so a nudge cannot push
past 0 or 1 and means the same thing at p=0.5 as at p=0.9. Every deviation is
attributable to a named axis and a signed nudge, stored as
challenger_adjustments, and the total is capped at 0.45 log-odds — a lean on a
real signal, never a re-forecast. Only mechanically obvious stat/axis
relationships are mapped; a speculative mapping would be the same guessing this
layer exists to replace.
IDENTICAL WHERE THERE IS NO SIGNAL, by construction. An unremarkable player, a
thin sample, an unmapped stat or a missing classification all return the
champion's probability byte-for-byte with an empty adjustment list and a stated
reason. The experiment therefore differs only where archetype-awareness could
possibly help or hurt, with no dilution from rows the treatment never touched.
Induced on real players. Judge home runs over: 0.42 -> 0.447, via BOMBER +0.22
and WHIFF RISK -0.11 — two real opposing signals netting positive. The same prop
under mirrors it exactly to -0.027. Judge strikeouts: delta exactly 0, because
WHIFF RISK and GRINDER cancel — an honest "no lean" with both signals still
recorded. Skubal strikeouts over: 0.60 -> 0.702 via WHIFF, TRAPDOOR and CANNON
all aligned; his hits-allowed goes the other way, 0.50 -> 0.392, because a
strikeout arm makes hits less likely. Josh Bell and a 12-PA sample are
untouched.
Isolation is structural: adjust() is pure, the champion field is read and never
written, the served snapshot payload is still the untouched champion object, and
a challenger failure is caught so it can never break the pipeline it is measured
inside. Statcast aggregates load once per snapshot run rather than per prop, so
grade-time I/O stays at zero.
Migration 034. Tests 3634 passed / 295 suites, web build exit 0.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01VCNgGSt5qvcLxaeQqa7Zpj
builtbykev
f2da9dd7e8
